Upload bench_eval_code/bench_utils.py with huggingface_hub

bench_eval_code/bench_utils.py

@@ -0,0 +1,98 @@
+"""
+Shared utilities for CPI-Bench evaluation scripts.
+Includes: API key pool, image helpers, retry-wrapped VLM caller.
+"""
+
+import base64
+import io
+import threading
+import time
+
+from PIL import Image
+from openai import OpenAI
+
+MAX_RETRIES = 3
+RETRY_DELAY_BASE = 2  # exponential backoff base (seconds)
+
+
+class ApiKeyPool:
+    """Thread-safe round-robin API key pool (supports multiple keys for higher QPS)."""
+
+    def __init__(self, keys: list):
+        if not keys:
+            raise ValueError("At least one API key must be provided.")
+        self._keys = list(keys)
+        self._index = 0
+        self._lock = threading.Lock()
+
+    def next_key(self) -> str:
+        with self._lock:
+            key = self._keys[self._index % len(self._keys)]
+            self._index += 1
+            return key
+
+
+def pil_to_base64(img: Image.Image, fmt: str = "PNG") -> str:
+    """Encode a PIL image (already RGB) to base64 string."""
+    if img.mode == "RGBA":
+        # flatten transparency onto white background
+        white_bg = Image.new("RGB", img.size, (255, 255, 255))
+        white_bg.paste(img, mask=img.split()[3])
+        img = white_bg
+    elif img.mode != "RGB":
+        img = img.convert("RGB")
+    buffer = io.BytesIO()
+    img.save(buffer, format=fmt)
+    return base64.b64encode(buffer.getvalue()).decode("utf-8")
+
+
+def load_local_image(path: str) -> Image.Image:
+    """Load a locally-saved result image (produced by the model being evaluated)."""
+    img = Image.open(path)
+    img.load()
+    if img.mode == "RGBA":
+        white_bg = Image.new("RGB", img.size, (255, 255, 255))
+        white_bg.paste(img, mask=img.split()[3])
+        return white_bg
+    return img.convert("RGB")
+
+
+def call_vlm_with_retries(
+    content_parts: list,
+    api_key_pool: ApiKeyPool,
+    base_url: str,
+    model: str,
+    max_tokens: int = 8192,
+    temperature: float = 0.1,
+    top_p: float = 0.95,
+    extra_body: dict | None = None,
+    tag: str = "",
+) -> str:
+    """
+    Call an OpenAI-compatible chat completions endpoint with retries.
+    Returns the raw text content, or a string starting with 'Error' on failure.
+    """
+    last_error = None
+    for attempt in range(1, MAX_RETRIES + 1):
+        api_key = api_key_pool.next_key()
+        client = OpenAI(api_key=api_key, base_url=base_url)
+        try:
+            kwargs = dict(
+                model=model,
+                messages=[{"role": "user", "content": content_parts}],
+                timeout=120,
+                max_tokens=max_tokens,
+                temperature=temperature,
+                top_p=top_p,
+            )
+            if extra_body:
+                kwargs["extra_body"] = extra_body
+            response = client.chat.completions.create(**kwargs)
+            return response.choices[0].message.content
+        except Exception as e:
+            last_error = e
+            if attempt < MAX_RETRIES:
+                delay = RETRY_DELAY_BASE * (2 ** (attempt - 1))
+                print(f"[{tag}] Attempt {attempt} failed: {e}. Retrying in {delay}s...")
+                time.sleep(delay)
+    return f"Error after {MAX_RETRIES} retries: {last_error}"